面向火灾的多尺度目标检测算法

A Multi-Scale Object Detection Algorithm for Fires

传统的火灾检测依靠各种传感器或者一些传统算法,只适用于特定场景,鲁棒性差,检测精度低,定位不准,并且检测时间长,不能满足实时性的要求。针对以上问题,提出一种基于YOLO V3的多尺度目标检测算法,自动提取火焰特征并实现多尺度的火焰检测。首先对YOLO V3的结构进行改进,将YOLO V3原先的特征提取网络Darknet-53替换为DenseNet,削弱网络加深带来的梯度消失问题,并且能够增强特征复用,提高网络对低层特征的学习,接着,为了消除感受野内像素作用的高斯分布,将DenseNet中的下采样改为空洞卷积。最后,优化锚框的定位方式,并根据新的定位方式修改损失函数,使网络对目标的定位更加准确。实验结果表明,改进的算法准确率和召回率为90%,85%。检测速度可以达到31帧/s,能够满足火灾检测对准确率和实时性的要求。

Traditional fire detection relies on various sensors or some traditional algorithms.These methods are only applicable to specific scenarios,with poor robustness,low detection accuracy,inaccurate positioning,and a long detection time,which cannot meet real-time requirements.For the above problems,a YOLO V3-based multi-scale object detection algorithm is proposed to automatically extract flame features and realize multi-scale flame detection.Firstly,the YOLOV3 structure was improved to replace original YOLO V3 feature extraction network Darknet-53 with DenseNet and weaken the gradient extinction problem caused by network deepening,which can enhance the feature taking and improve the network learning of low-level features.Then,in order to eliminate the Gaussian distribution of pixels in the receptive field,the downsampling in DenseNet was changed to dilated convolution.Finally,the positioning mode of anchor box was optimized,and the loss function was modified according to new positioning mode,to make the network more accurate in locating targets.Experimental results show that the improved algorithm accuracy and recall rate are 90%and 85%respectively.The detection speed can reach 31 frames/s,which can meet the accuracy and real-time requirements of fire detection.